[LamaSu]

Why

The 9/29 install-path audit found that our public install instructions don't work: agents that read llms.txt, pcc.json or the landing page are told to run commands that fail. The ADK and hackathon track depends on agents being able to connect. The steward's list for this lane (from the audit and rehearsal R0) covers the stale docs outside the agent pack: install.html, root AGENTS.md, MCP_INSTALL.md and the Claude Code quickstart. The agent pack and the pcc skill belong to adk (#452).

Advertised	What happens	This PR
pip install pcc-node	Fails today: 0.1.0 is yanked (it ships the N66 shell executor), and 0.1.1 isn't on PyPI yet.	pip install "pcc-node[crypto]>=0.1.1", the form #452 uses. 0.1.1 is due by 10-01 (board ADK item 1).
npx -y @pcc/mcp-server, npm install -g @pcc/operator-agent-runtime, pnpm add @pcc/operator-agent-runtime	npm E404. No @pcc package is published, and the operator runtime isn't in the repo either.	Removed. Replaced with the hosted MCP server and the job-offers API calls that work today.
/skills install https://capability.network/skills/pcc.md	No such command. In Claude Code 2.1.285, /skills only lists skills.	mkdir -p ~/.claude/skills/pcc, then curl the skill into it.

Merge order

The pip lines assume 0.1.1 is on PyPI. Merge this after item 1 publishes it. Nothing merges before Oct 5 anyway, and the target is 10-01. If 0.1.1 slips past then, hold this PR: until then, the floored command fails with "No matching distribution found for pcc-node>=0.1.1".

The hosted MCP lines also need a prod setting. Today https://capability.network/mcp answers initialize, then fails tools/list with -32600 PCC_API_BASE_URL is not set — the MCP proxy is disabled. #271's fail-closed guard shipped, and the variable was never set on prod, so every MCP client gets zero tools. That is a live hole, and it is on master already: llms.txt, the skill and the MCP Registry listing all point there. The fix is one Railway variable, PCC_API_BASE_URL=https://capability.network, and it's the operator's. This PR's claude mcp add lines are right once it's set.

What changes

• pcc-node: these files install with pip install "pcc-node[crypto]>=0.1.1":

  • llms.txt and pcc.json (onboarding.one_command);
  • PCC-NETWORK.md, README.md and AGENTS.md;
  • docs/AGENT_INTEGRATION.md;
  • the landing page (its three CTAs, the terminal demo and step 1).

  The floor keeps everyone off the yanked 0.1.0. The crypto extra is what the node needs to sign. Double quotes make the command work in bash, zsh, PowerShell and cmd.exe.

• Landing page: its three "GitHub" links pointed back to capability.network. They now point to the repo.

• @pcc on npm:

  • PCC-NETWORK.md and the quickstart no longer install @pcc/operator-agent-runtime.
  • pcc.json marks that runtime not built.
  • The quickstart shows what works now: GET /api/job-offers/open?capabilityType=… (public), then POST /api/job-offers/:id/claim with a key and the kernelId body the route requires.

• docs/MCP_INSTALL.md is now a short pointer to the one public copy, apps/dashboard/public/MCP_INSTALL.md, which fix(docs,gateway): install calls to action that work, and no local paths (N26) #414 rewrites and the gateway serves at /MCP_INSTALL.md. It was a second copy that had drifted and still said npx -y @pcc/mcp-server (and a personal C:/Users/… path). The pointer keeps the one-line hosted-MCP command.

• /skills install: the /start page (start_here in pcc.json) and the quickstart now show the download that works.

  • The quickstart's manual install ran curl -o into a directory before creating it; the order is fixed.
  • The quickstart offers claude mcp add --transport http pcc https://capability.network/mcp as the alternative that needs no install.

• install.html:

  • It says "250+ tools", not "230+" (the pack has 253).
  • It drops the "~207 KB" size (the pack is 265 KB).
  • The hand-off prompt now provisions the key before the template call that needs it (/api/sdk/adapter-template/* returns 401 without one).
  • It says @pcc/kernel is not on npm, so the agent works in a clone.
  • It names MachineAdapter's real methods, getStatus, execute, getProgress, onEvidence and dispose, which the served templates also implement. It used to list connect, getHealth and executeJob, which exist nowhere.

• Guard test apps/dashboard/src/__tests__/agent-facing-install-honesty.test.ts (8 tests) pins the install matcher with a table of forms it must flag (including python -m pip, ==0.1.0, a floor without the extra, a line continuation and uvx) and forms it must pass. It checks that:

  • every pcc-node install in the agent-facing files and in public/ names 0.1.1 or later with the crypto extra, the same rule as fix(agent-pack): only true statements, attempt reporting, #427 folded in, the operator's thesis (ADK item 2: N75, N76, rehearsal R0, contract v1) #452's agent-pack-truth.test.ts;
  • no agent-facing file installs an @pcc package from npm;
  • nothing says /skills install;
  • every claude mcp add targets the gateway's remote endpoint.

  pcc.json is in the checked set, so a later sync from LamaSu/pcc-network-kit that brings back an old command fails CI.

Verification

• Guard test: against master's files, 5 of the 7 file checks fail, and each failure names a site fixed here. All 8 tests pass on this branch.
• Other tests:
  • OnboardLandingPage.test.tsx plus the guard: 19 of 19 pass.
  • well-known-404.test.ts (gateway, covers /start): 8 of 8 pass.
• Parse checks:
  • LandingPage.tsx: 0 TypeScript syntax diagnostics.
  • install.html's inline script passes node --check.
• /start: rendered through Fastify inject, it returns 200 and shows the two-line install.
• Quickstart install: I ran it in a throwaway HOME. It exited 0 and wrote a 16,053-byte SKILL.md with valid frontmatter.
• Prod probes today (keyless, read-only):
  • GET /api/job-offers/open?capabilityType=manufacturing.fdm returns 200.
  • POST /mcp initialize returns 200, but tools/list returns the PCC_API_BASE_URL error above. initialize alone doesn't prove the server works.
• The claim route needs {"kernelId": …} in its JSON body, or it answers 400 missing_field (job-offers.ts:268-270). The quickstart's example sends it, and PCC-NETWORK.md says so.
• PyPI today: only 0.1.0, both files yanked. pip download "pcc-node>=0.1.1" finds no distribution, which is why the merge-order note above exists.
• Trial merges:
  • fix(agent-pack): only true statements, attempt reporting, #427 folded in, the operator's thesis (ADK item 2: N75, N76, rehearsal R0, contract v1) #452, fix(docs,gateway): install calls to action that work, and no local paths (N26) #414, fix(copy): align published copy with the public-beta status and add a claims check #364, docs(site): lead published descriptions with the product thesis (stacked on #364) #386 and fix(agent-package): drop node-local pcc_generate_ui from the public package #427 each merge cleanly with this branch.
  • In the fix(agent-pack): only true statements, attempt reporting, #427 folded in, the operator's thesis (ADK item 2: N75, N76, rehearsal R0, contract v1) #452 merge, every pcc-node line in the skill copies and CLAUDE.md is floored.
  • The docs(site): lead published descriptions with the product thesis (stacked on #364) #386 merge keeps this PR's floored llms.txt line.

Not in this PR

• adk (fix(agent-pack): only true statements, attempt reporting, #427 folded in, the operator's thesis (ADK item 2: N75, N76, rehearsal R0, contract v1) #452): the pcc skill (all three copies) and CLAUDE.md. fix(agent-pack): only true statements, attempt reporting, #427 folded in, the operator's thesis (ADK item 2: N75, N76, rehearsal R0, contract v1) #452 already floors their pip lines and fixes CLAUDE.md's counts. fix(agent-pack): only true statements, attempt reporting, #427 folded in, the operator's thesis (ADK item 2: N75, N76, rehearsal R0, contract v1) #452 uses single quotes, 'pcc-node[crypto]>=0.1.1', which break in cmd.exe. Both tests accept either quote style.
• adk (item 1): apps/dashboard/public/install.sh installs pcc-node from unpinned master (git+…#subdirectory=packages/pcc-node). Every install.html user therefore gets master's N66 shell executor until fix(pcc-node): remove the relay executor that ran commands in a shell (N66) #442 merges. Item 1's pin, and then the floor, should cover this script.
• pcc-launch: apps/dashboard/public/whitepaper.md still has 3 unfloored pip install pcc-node lines. It is allowlisted in the test, and a ratchet test fails once that file is fixed, so the entry gets removed.
• adk (N26 and the pack):
  • the package READMEs, including packages/mcp-server/README.md;
  • EXTENDING_PCC.md and digital-kernel-marketplace.md;
  • onboard-kit/AGENT_INSTRUCTIONS.md (private @pcc/* imports);
  • pcc-node's own strings;
  • install.html's hand-off flow as a whole. It asks for a TypeScript adapter, while the kernel daemon operators run is Python pcc-node. Items 5 and 7 replace it.
• fix(copy): align published copy with the public-beta status and add a claims check #364 (launch): it already removes the dead npmjs.com links on /start, so this PR leaves those lines alone.
• R43(b) facts generator (PX-16a), separate PR: hand-kept counts, such as README's "230+", llms.txt's "254" and /start's "63 tools" (right today).
• R43(c): AGENTS.md's pcc_setup_* workflow names are the stdio server's, which is correct for the config shown there. The hosted /mcp uses the agent package's names (setup_detect, …). They converge when the stdio server generates from the pin (feat(adk): @pcc/adk skeleton with a pinned agent package and a clean external install check (R4 D1) #411 follow-up).

Ledger rows: R36, R43.

🤖 Generated with Claude Code